Combined Convolutional and LSTM Recurrent Neural Networks for Internal Defect Detection of Arc Magnets Under Strong Noises and Variable Object Types

This paper focuses on developing the acoustic-based detection of arc magnet internal defects under different noises and object types. The major challenge in such a detection case is that the traditional approaches mainly rely on expert knowledge and experiential features, which results in a poor generalization performance of the algorithm for new types of detection objects and is susceptible to noise interference. This work presents a novel detection framework based on an end-to-end one-dimensional (1D) convolutional long short-term memory (LSTM) model, where both the spatial and temporal features of the measured acoustic signals are extracted and then jointed for determining the internal defects of arc magnets. In addition, the LSTM layers are employed behind the 1D convolutional neural network, which makes the number of time steps in the LSTM layers for the temporal feature extraction is much smaller than the length of the input segments, thus the computational complexity of the LSTM layers can be highly reduced. Experimental results show that our method is superior to existed methods in the detection accuracy for the internal defects of arc magnets, and the diagnosis time per a single arc magnet is controlled at the millisecond, making it appropriate for real-time applications. Furthermore, the robustness of the proposed framework is validated through experiments on different signal-to-ratios and multiple object types of arc magnets.